Lang.NET 3 Years OnMon 27 April 2009 by Jim Purbrick
It was incredibly satisfying to be able to go back to Lang.NET 3 years on to be able to say that we actually managed to make all the crazy plans we had for Mono in 2006 work. My talk is now online here. Lots of people hadn’t seen the 2006 talk and were blown away with us adding support for continuations on the CLR and enough new stuff to keep those that had seen the first talk interested. In particular the anecdote about flywheel exploits for the Mono scheduler got a laugh from Anders Hejlsberg.
The other highlights on Tuesday were Mads Torgeson’s talk, which gave some nice insights in to the process that led up to the final C# 4.0 Dynamic design. Mads did a good job capturing some of the concerns that static advocates have with dynamic languages. Gilad Bracha’s talk on Newspeak and the Hopscotch IDE showed what’s possible with a really dynamic language: extending the IDE and debugging it from inside itself. The access that newspeak has to it’s call stack was particularly interesting: it would be much easier to move a newspeak stack around than build the infrastructure we needed to move a CLI stack. I spoke to Gilad about his experiences building newspeak on Sqeak afterwards. His impression is that Squeak is runtime mostly used for education and has no support for security or running untrusted code. Having spent a decade seeing the various security problems with bytecode verification in Java, Gilad isn’t convinced that a bytecode level sandbox is the right approach, but although he has some ideas for object capability based security for the post squeak Newspeak implementation it seems a long way off.
The highlight on Wednesday was definitely Lars Bak’s talk on V8, which I had missed last year at JAOO. The mechanism for discovering emergent types in dynamic languages to allow indexed lookup instead of hash table lookup is a very nice hack. Lar’s super competitive heckling of everyone at Microsoft for the rest of the day was also fun. Other highlights included Erik Meijer throwing coins around while talking about mathematical duality and Amanda Lauter and Luke Hoban talking about F#. Potentially relevant to Linden Lab if we get to the point where we can support debugging scripts in Second Life was Herman Venter’s talk on the Common Compiler Infrastructure a newly open sourced library which can allow CLI assembly rewriting while preserving debug information.
The Meta Introduction to Domain Specific Languages was a really good opening for the DevCon, but the talk most relevant to Second Life was Paul Cowan talking about DSLs in the Horn Package Manager. Paul talked about creating DSLs by extending Boo, something that should be possible when we get to the point where Second Life can accept arbitrary CLI assemblies. I got a first Boo script compiling against the Second Life assemblies at the DevCon and have some plans to experiment building a DSL for virtual ecosystems in Second Life over the next few weeks. Supporting C# as a general purpose language and Boo as a meta language for building DSLs in Second Life would be an excellent combination.
A lot of the talks at the DevCon seemed to involve doing terrifying language somersaults to create more natural DSLs that run on the JVM or CLI which then can’t be easily debugged due to the huge chained expressions or nested types that the languages boil down to. There seems to be a large opportunity for disaster here when these DSLs need to be extended or maintained in 10 years time after the original authors have moved on. Although Martin deflected a lot of these questions by saying you can get bad frameworks as easily as bad DSLs, I feel a lot more comfortable unpicking a bad framework or wrapping a library than trying to fix or extend a broken language (maybe several years reimplementing LSL’s quirks did that).
The ultimate promise of DSLs was hinted at by Magnus Christerson’s talk which demoed Intentional’s amazing Domain Workbench. Instead of building DSLs on top of mainstream runtimes, the Domain Workbench models the domain and can then project the model as code or using domain specific projections like interactive circuit diagrams that can be manipulated and debugged interactively. Magnus started his talk saying that we used to have to enter programs on punch cards that evolved in to sequential programs and that we don’t need to do that any more. If the Domain Workbench’s promise is fulfilled it could change the way we develop software forever.